Track auxiliary V2 LLM usage in session lifetime totals #35781

Summary

V2 session lifetime usage currently counts coding-agent steps, but omits LLM calls made for session title generation and compaction. Both paths consume an LLM stream directly and discard settled usage, so their tokens and cost never reach Session.Info or session.usage.updated.

Current behavior

• packages/core/src/session/title.ts consumes title-generation stream events but retains only text and provider failure state.
• packages/core/src/session/compaction.ts consumes compaction stream events but retains only text and provider failure state.
• packages/core/src/session/projector.ts increments lifetime usage only from usage-bearing session.step.ended and session.step.failed events.
• Forked sessions correctly start their own lifetime totals at zero; copied transcript tokens remain available separately for context-window display.

Expected behavior

• Capture normalized token usage for title-generation and compaction calls.
• Calculate cost through the same typed pricing/billing path used by normal model steps.
• Attribute those calls once to the owning session lifetime totals, including settled failed calls when the provider reports usage.
• Persist an irreducible durable usage fact so replay and retries cannot lose or double-count spend; do not rely on an untracked direct SessionTable mutation.
• Continue publishing absolute session.usage.updated snapshots after the aggregate is projected.
• Do not represent auxiliary calls as assistant steps or include them in context-window usage.

Implementation notes

A dedicated durable usage event may be cleaner than adding unrelated billing fields to session.renamed and compaction completion events. The event should carry a stable source identity or otherwise preserve exactly-once accounting under retries.

Compaction currently receives a route-level model without catalog pricing, so its input may need the resolved model cost information or a shared billing service.

Tests

• Successful title generation increments lifetime tokens and cost once.
• Successful automatic and manual compaction increment lifetime tokens and cost once.
• Usage-bearing failures are counted once.
• Retry/replay does not duplicate usage.
• Auxiliary usage does not change context-window percentage.

Related

• Add typed Copilot billing usage to the V2 runtime #35765 tracks the typed provider-reported billing contract, including Copilot billing.
• v2 SessionRunner publishes Step.Ended.cost: 0; Copilot refresh drops Anthropic cache-write pricing in token fallback #30706 tracks broader V2 cost-accounting history and Copilot fallback pricing.